Agentic + AI Night: Multi-Agent Governance and Security

Join us for another Agentic + AI Night: Multi-agent Governance and Security on Tuesday, September 1 from 3:30pm - 9:00pm PT in SF for a multi-track deep dive into agentic architectures, security governance, and AI observability using Omnigent, an open-source meta-harness for coordinating agents such as Claude Code, Codex and custom agents in a shared workflow. This meetup is built for engineers, security practitioners, ML/AI developers, and AI startup founders who are building or deploying autonomous agents and want to go deeper into the tech. We’ll dive into real-world patterns, failure modes, and governance models for building a reliable meta-harness layer that manages agent execution, sandboxing, state, and runtime policies across multiple agents and model providers. Whether you’re at an early-stage startup or an established enterprise, if you care about deploying agents in your organization’s workflows safely with clear controls and observability, this meetup is for you. You’ll hear from the following leaders in the AI space: Databricks, OpenAI, Cursor, OpenHands, Ollama, and Modal! 💡 Why You Should Attend See Real Security Architectures: Learn how teams design the meta-harness layer on top of isolated sandboxes, separating model reasoning from privileged execution environments to prevent agent runaway. Observe & Audit Agent Decisions: Go beyond standard logs to structured traces, evals, and audit trails that track agent choices, enforce security policies, and monitor cost budgets. Get Hands-On with Omnigent: Watch live demos of Omnigent—orchestrating cross-vendor agents (OpenAI, Codex, custom models) in adversarial, verification-driven workflows. Connect with Other Builders: Meet engineers, founders, and security practitioners working on agent autonomy, swap meta-harness design patterns, and find potential collaborators. AGENDA 3:30pm: Registration & Mingling 5:00pm: Keynote with Kasey Uhlenhuth, Director, Product Management, Databricks 5:45pm: Breakout #1 (2 tracks)[TERRA] Meta Harnesses: The new layer for secure, collaborative, multi-agent coding with Bryan Qiu, Databricks [MER] Building Security Agents with Travis McPeak, Cursor 6:30pm: Breakout #2 (2 tracks)[TERRA] From Vulnerability to Verified Fix: Governing Cyber Agent Teams Across OpenAI and Databricks with Tom Petrick and Saptarshi Banerjeej, OpenAI [MER] Coding Agents with Open Models: Best Practices, Jeffrey Morgan, Ollama 7:15pm: Breakout #3 (2 tracks)[TERRA] Self-Healing Software with an Agentic Fleet: How OpenHands Uses OpenHands to Build and Secure OpenHands with Cliff Yang and Saurya Velagapudi, OpenHands [MER] The Time-Travel Filesystem for Agents with Erik Dunteman, Modal 7:45pm: Reception 9:00pm: Goodnight Session Descriptions: Meta Harnesses: The new layer for secure, collaborative, multi-agent coding with Bryan Qiu, DatabricksCoding agents have become remarkably capable, and many software developers now depend on them for a large part of their day-to-day productivity. But the experience around them is still fragmented and hard to manage: for example, every harness and agent framework has a different interface, spending and security controls vary, sessions are often tied to a developer’s laptop, and teammates usually can’t see or steer them. In this talk, we demonstrate how Omnigent addresses these limitations by giving developers one place to use different coding agents and harnesses, build custom agents, run them remotely, collaborate on live sessions, and access them from any device. We’ll also share what we learned deploying Omnigent to thousands of engineers at Databricks. Building Security Agents with Travis McPeak, CursorSecurity agents have enabled security teams to scale impact beyond what anybody would have considered possible a few years ago. Every change can now be diligently reviewed. In this session Travis will share what his team learned implementing security agents at scale at Cursor and the positive impact the agents have had. From Vulnerability to Verified Fix: Governing Cyber Agent Teams Across OpenAI and Databricks with Tom Petrick and Saptarshi Banerjeej, OpenAIMost security-agent demos end when a model generates a patch. This session asks the harder question: How do we prove that the patch actually works? Using a synthetic vulnerable application, we’ll demonstrate a governed cyber-remediation workflow spanning OpenAI and Databricks. An OpenAI-powered agent identifies a vulnerability and assembles the supporting evidence. An independent cyber agent deployed through a Databricks App then attempts to reproduce the issue. Once the finding is confirmed and a person authorizes remediation, Codex creates a candidate patch and regression tests. However, the agent that writes the patch does not get to certify it. The Databricks-hosted verifier attacks the patched application, rejects an initially plausible but incomplete fix, and supplies the counterexample needed to produce a stronger patch. Omnigent orchestrates the agents, evidence handoffs, and human approvals, while MLflow records the complete trace and evaluation history. Attendees will see a practical pattern for moving from autonomous patch generation to evidence-backed, human-governed security remediation. Coding Agents with Open Models: Best Practices, Jeffrey Morgan, OllamaOpen models now power a meaningful share of coding agent usage. This talk covers how to choose an open model for coding agents, where to deploy them (local, self-hosted, or cloud), and the privacy and cost tradeoffs of each. It also covers using leading multi-agent frameworks such as Omnigent with open models, including best practices for keeping open model-powered agents secure with sandboxing, observability, and other safeguards. Self-Healing Software with an Agentic Fleet: How OpenHands Uses OpenHands to Build and Secure OpenHands with Cliff Yang and Saurya Velagapudi, OpenHandsWhat happens when you turn a fleet of coding agents loose on your own codebase — and trust them to keep it secure? At OpenHands, we run our agents in production against our own product: automatically detecting and remediating CVE exposure, triaging bugs, and shipping fixes with minimal human intervention. In this session, we'll show how we orchestrate this agentic fleet safely, and how the observability dashboards we've built let us debug, govern, and trust multi-agent workflows at scale. We'll share real battle stories from production — where agents succeeded, where they failed spectacularly, and the guardrails and observability patterns that made the difference. You'll leave with a practical blueprint for deploying self-healing automation on your own projects, and a clear-eyed view of what agentic security looks like in practice today. The Time-Travel Filesystem for Agents with Erik Dunteman, ModalIn this talk, we'll explore how append-only-logs can improve your agent systems with perfect observability, durability, and even time travel! We'll demo Overeasy, an experimental filesystem at Modal, to demonstrate the power of backing agent state with the almighty append-only-log.